Prediction And Comparison Of Spatial Distribution Of Soil Heavy Metals Based On Different Factors And Models

Soil is one of the main natural resources on which human beings depend and also an important part of human ecological environment.In recent years,soil heavy metal pollution has caused serious consequences.It has caused a shortage of agricultural land resources.This poses a great threat to the health of the people and has a great adverse impact on agricultural development.Therefore,carrying out the prevention and control of heavy metal pollution in soil has become an urgent need of the current soil environmental protection work.Accurately grasping the characteristics of spatial variability of soil heavy metals is the key and prerequisite for the prevention and control of heavy metal pollution in soil.Under the current background of increasingly severe soil environmental problems,the optimal utilization of soil resources and the protection and management of soil environment urgently need to accurately grasp the contents of heavy metals.However,it is not feasible to take a large number of samples in reality because of the limitation of human and material resources.Therefore,it is of great significance to use scientific methods to predict regional soil properties with fewer samples.At present,such methods as geostatistics,fuzzy clustering and neural network have been applied to the studies of spatial heterogeneity of heavy metals in soil to explore and improve the spatial prediction techniques of heavy metals in soil.However,there are few comparative studies on the prediction accuracy of different models.In view of this,it is of important scientific and practical significance to explore and improve the spatial prediction techniques and to compare and optimize the accuracy of different prediction models.This paper took Jintan District,Changzhou City,Jiangsu Province as the study area,relying on national natural science projects(41771243),special research projects of the Ministry of Land and Resources(20151001-03)and Jiangsu Provincial Land and Resources Technology Project(201406).180 samples were selected for soil sample collection and testing,and further supplement data were collected from other sources such as land use data and basic geographic information data in the study area.Considering the modeling elements from source-sink relationships,spatial differentiation,and the combination of the two aspects,using linear regression,geostatistical interpolation,neural networks,geographically weighted regression and other kinds of models,this paper constructed spatial distribution models of Cd,Pb,Cr,Cu,and Zn in the soil of the study area.Based on this,comprehensive optimization methods were used to predict soil heavy metal contents in the study area.According to the analysis and the comparison,the following conclusions are drawn:(1)Compared with the LUR model,the goodness of fit is better and the prediction accuracy is generally higher.Therefore,the factors of source-sink relationship should be considered when constructing the prediction model of spatial distribution of heavy metals in soil.(2)By comparing the goodness of fit and prediction accuracy between LUR-S model and BP-S model and between OK model and BP-K model,it can be found that BP-S model and BP-K model are better than LUR and OK models,which shows that the neural network model is more accurate because of the consideration of the spatial relationship between spatial clustering and non-linear relationships among variables.(3)Predicting the distribution of heavy metal content in the whole study area by comparing different models,it can be found that the BP-SK and GWR models,which are modeled by considering the source-sink relationship and spatial differentiation,show better prediction results by combining the key information of models based on source-sink relationship and models based on spatial differentiation,and the prediction results are most similar to the reference distribution.(4)According to the prediction accuracy of heavy metals in soil in the study area,it shows that BP-SK,BP-S and BP-K neural network models have higher prediction accuracy for Cd contents.BP-K and OK models which are based on the spatial differentiation have higher accuracy of prediction of Pb contents.GWR and BP-K models have higher accuracy of prediction of Cr contents.BP-S,BP-K and BP-SK neural network models have higher prediction accuracy of Cu contents.The LUR-S model has higher prediction accuracy for Zn contents.(5)Generally speaking,BP-S model is the best model for predicting soil heavy metal content in the hilly region in the western study area,and BP-SK and BP-S are the best models for the urban and rural areas in the central study area.For the plain agriculture region in the eastern sduty area,the optimal model for the area is the BP-SK and GWR models.(6)In some prediction results,the model prediction results based on the dual consideration of source-sink relationship are similar with those based on the source-sink relationship.This may be due to the fact that the model variables based on the source-sink relationship may include spatially disparate information and that was implicitly added to the model based on source-sink relationships.